Elastic Learned Sparse Encoder

Overview

Elastic Learned Sparse Encoder (ELSER) is a learned sparse encoding model developed by Elastic that brings semantic search capabilities to Elasticsearch while maintaining the efficiency and explainability of traditional keyword search.

Key Innovation

ELSER combines the best of both worlds:

• Semantic Understanding: Neural network-based semantic comprehension
• Sparse Representation: Efficient sparse vectors for fast retrieval
• Term Expansion: Automatic expansion with related terms
• Native Integration: Built directly into Elasticsearch

How It Works

Architecture

ELSER uses a learned sparse encoding approach:

1. Term Expansion: Expands queries and documents with semantically related terms
2. Sparse Vectors: Generates sparse vector representations
3. Efficient Storage: Leverages Elasticsearch's inverted index
4. Fast Retrieval: Uses BM25-like retrieval with semantic enhancement

Advantages of Sparse Encoding

• Storage efficient compared to dense vectors
• Faster retrieval than approximate nearest neighbor search
• Explainable results (can see which terms matched)
• No need for separate vector database infrastructure

Performance

ELSER achieves strong retrieval performance:

• Semantic understanding comparable to dense embeddings
• Faster query execution than dense vector search
• Lower storage requirements than dense vectors
• Better zero-shot performance on domain-specific queries

Features

• Zero-Shot Learning: Works without domain-specific training
• Multilingual Support: Handles multiple languages
• Explainability: Clear visibility into why documents matched
• Hybrid Search: Can be combined with traditional BM25 search
• Native Integration: No external embedding service required
• Automatic Deployment: Easy setup within Elasticsearch

Use Cases

• Enterprise search applications
• Document retrieval systems
• Question answering platforms
• Knowledge base search
• E-commerce product search
• Legal document discovery

Comparison with Dense Vectors

ELSER Advantages

• Faster retrieval speed
• Lower storage costs
• Explainable results
• Better integration with existing Elasticsearch features

Dense Vector Advantages

• Higher semantic precision in some cases
• Better for cross-modal search (image-text)

Integration

Elasticsearch Setup

ELSER can be deployed directly in Elasticsearch:

PUT _ml/trained_models/.elser_model_2
{
  "input": {
    "field_names": ["text_field"]
  }
}

Ingestion Pipeline

Automatic inference during document indexing:

PUT _ingest/pipeline/elser-ingest
{
  "processors": [
    {
      "inference": {
        "model_id": ".elser_model_2",
        "input_output": [
          {
            "input_field": "content",
            "output_field": "content_embedding"
          }
        ]
      }
    }
  ]
}

Search Query

GET my-index/_search
{
  "query": {
    "text_expansion": {
      "content_embedding": {
        "model_id": ".elser_model_2",
        "model_text": "How to install security patches?"
      }
    }
  }
}

Model Versions

• ELSER v1: Initial release
• ELSER v2: Improved performance and accuracy
• Regular updates with enhanced capabilities

Performance Characteristics

• Query latency: Sub-100ms typical
• Storage: ~100 tokens per document on average
• Recall: Competitive with dense vector approaches
• Throughput: High queries per second

Best Practices

• Use ELSER for text-heavy enterprise search
• Combine with traditional BM25 for hybrid search
• Monitor resource usage during inference
• Use appropriate field mappings for optimal performance
• Consider document length when planning capacity

Advantages for Production

• No external embedding service required
• Unified infrastructure (no separate vector database)
• Leverages Elasticsearch scaling and reliability
• Familiar query syntax and operations
• Built-in monitoring and management

Pricing

Included with Elasticsearch, available on:

• Elastic Cloud (managed service)
• Self-managed Elasticsearch deployments
• Pricing based on Elasticsearch licensing tiers